Hierarchical Multi-Grained Reasoning for Object Concept Learning

Yuxuan Wang, Muli Yang, Yihang Zhu, Aming WU, Cheng Deng

Keywords: Hierarchical Multi-Grained Reasoning for Object Concept Learning

Abstract:

Human beings can easily understand object concepts involving attributes and affordances. Recently, to simulate this ability, Object Concept Learning (OCL) has been introduced as a new task to recognize attributes and affordances related to a given object.
OCL is essentially a many-to-many mapping problem: While an object may possess multiple different concepts, a concept can also belong to multiple different objects.
In this regard, the prevailing method of learning discriminative representation---which is effective in the single-mapping cases---often fails in OCL.
Inspired by the reasoning mechanism of human beings, in this paper, we propose Hierarchical Multi-Grained Reasoning (HGR) for OCL, aiming to infer object-related concepts from coarse-to-fine and counterfactual grains.
Specifically, we first propose a coarse-to-fine hierarchical reasoning module that exploits multi-step learnable prompts to progressively localize object-relevant concept information. Subsequently, multiple counterfactual samples are selected to strengthen the relations between objects and concepts, which further improves the reasoning performance. In the experiments, our method is evaluated on multiple benchmarks. Significant performance gains and extensive visualization analysis demonstrate the superiorities of our method.

Primary Area: unsupervised, self-supervised, semi-supervised, and supervised representation learning

Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.

Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide.

Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors’ identity.

No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.

Submission Number: 4520